reference-video-director
Creates production-ready reference-driven video prompts from character, outfit, scene, relationship, story, and duration inputs. Use for requests such as 情侣短剧、情侣互动、男友视角、女友视角、情侣POV、恋爱POV、参考图视频、15秒/30秒剧情、couple POV, boyfriend POV, girlfriend POV, romantic short scene, reference video prompt, RISE火焰视频、点烟抛火柴连续镜头、燃烧文字、RISE fire video, burning-letter reveal, or when a rough scene needs natural dialogue, micro-expressions, continuity, and model-specific optimization.
Safety scan
A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.
Badge for your README
Markdown
[](/listings/sgyno09-source--reference-video-director)HTML
<a href="/listings/sgyno09-source--reference-video-director"><img src="/api/listings/sgyno09-source--reference-video-director/badge.svg" alt="findsafeskills safety rating" /></a>Install
Install instructions from the README
As written by the repo's author. findsafeskills has not verified these steps.
这个仓库本身就是一个 Skill 目录,核心入口文件是:
SKILL.md
兼容 Agent Skills / Codex Skill 风格的客户端,可直接使用仓库目录作为 Skill 来源。
What was scanned
The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.
SKILL.md · Manifest · 496 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- ---
- name: reference-video-director
- description: Creates production-ready reference-driven video prompts from character, outfit, scene, relationship, story, and duration inputs. Use for requests such as 情侣短剧、情侣互动、男友视角、女友视角、情侣POV、恋爱POV、参考图视频、15秒/30秒剧情、couple POV, boyfriend POV, girlfriend POV, romantic short scene, reference video prompt, RISE火焰视频、点烟抛火柴连续镜头、燃烧文字、RISE fire video, burning-letter reveal, or when a rough scene needs natural dialogue, micro-expressions, continuity, and model-specific optimization.
- ---
- # Reference Video Director
- ## Specialized fire-video workflow
- For RISE fire videos, lighting a cigarette and throwing a match, burning-letter reveals, or variants of this sequence, read [RISE Fire Video Prompt](skills/create-rise-fire-video/SKILL.md) and its [30-second baseline](skills/create-rise-fire-video/references/rise-30s.md). Follow that workflow for the full prompt and validation.
- In this mode, use its single continuous dolly-back shot, calm performance, mouth-held cigarette, one-match causality, delayed leg crossing, and unobstructed final composition. Do not apply the couple relationship defaults, changing eye-line requirements, recommended multi-shot counts, or couple dialogue/story engine below. Preserve the shared identity and outfit continuity rules. Treat RISE as scene geometry, not a prohibited text overlay. User-requested changes remain authoritative.
- For other scenes, continue with the general workflow below. Do not introduce smoking, fire, or burning letters into unrelated couple scenes.
- ## Purpose
- Turn a user's reference images and rough story idea into an executable short-form video prompt with stable identity, believable acting, natural dialogue, continuous blocking, realistic POV cinematography, and model-aware constraints.
- Default to the user's current language. If the user writes Chinese, answer in Chinese. If the user writes English, answer in English. If the user explicitly requests another language, follow that request.
- ## Trigger logic
- Activate when the user asks for one or more of the following:
- - reference-image-driven video prompting
- - couple / relationship POV
- - boyfriend POV or girlfriend POV
- - romantic daily-life short scenes
- - natural couple dialogue
- - fake anger, teasing, awkward comforting, playful banter
- - 10s / 15s / 30s / 60s narrative video prompts
- - timeline-based prompt writing
- - character identity locking
- - outfit or scene continuity
- - optimization for Seedance, Hailuo, Kling, Veo, Sora, or comparable video generators
- Common Chinese triggers include:
- `情侣短剧` `情侣互动` `恋爱短剧` `男友视角` `女友视角` `情侣POV` `恋爱POV` `女友假生气` `男友哄女友` `情侣聊天` `打情骂俏` `甜蜜互动` `参考图视频` `视频剧情提示词`.
- Common English triggers include:
- `couple POV` `boyfriend POV` `girlfriend POV` `romantic short film` `couple interaction` `relationship POV` `romantic dialogue` `realistic couple video` `reference video prompt`.
- Do not activate for unrelated general writing or purely technical video encoding tasks.
- ## Inputs
- Infer as many of these as available. Do not force the user to fill every field.
- - `{{Mixed 1}}`: primary character reference, usually the female lead
- - `{{Mixed 2}}`: outfit reference
- - `{{Mixed 3}}`: optional scene / environment reference
- - relationship
- - personality
- - story beat
- - target duration
- - target model
- - aspect ratio
- - audio requirements
- - camera POV
- - continuity constraints
- When information is missing, make conservative defaults that preserve the user's concept.
- ## Character consistency
- When a character reference is provided, treat it as the identity source of truth.
- Lock:
- - face shape
- - facial proportions
- - eyes
- - nose
- - lips
- - eyebrows
- - hairstyle
- - hair color
- - skin tone
- - perceived adult age
- - body proportions
- - overall recognizability
- Do not redesign the face unless the user explicitly asks.
- Prevent:
- - face swapping
- - identity drift
- - sudden makeup changes
- - hairstyle changes
- - age changes
- - unexplained body-proportion changes
- If the user provides multiple character references, clearly map each reference to one character.
- ## Outfit consistency
- When an outfit reference is supplied, lock visible clothing attributes:
- - garment type
- - color
- - material
- - straps
- - sleeves
- - neckline
- - lower garment
- - footwear
- - accessories
- Do not allow unexplained wardrobe changes within a continuous scene.
- ## Scene consistency
- When a scene reference is supplied, preserve the major spatial anchors:
- - room shape
- - bed / sofa / table
- - windows
- - curtains
- - mirrors
- - lamps
- - doors
- - major furniture placement
- - lighting direction
- - time of day
- Do not turn an ordinary room into a luxury hotel unless requested.
- ## Default relationship tone
- If the user does not specify otherwise, assume the characters are adult partners in a stable, familiar relationship.
- Favor:
- - familiarity
- - teasing
- - casual physical proximity
- - unfinished sentences
- - little pauses
- - playful complaints
- - shared jokes
- - believable reactions
- Avoid:
- - forced flirting
- - domineering romance tropes
- - constant intense eye contact
- - melodramatic confession language
- - scripted idol-drama behavior
- ## Dialogue rules
- Dialogue should be short, speakable, and slightly imperfect.
- Prefer:
- - short lines
- - interruptions
- - half-finished thoughts
- - small pauses
- - brief repetition
- - understated replies
- - reaction before response
- Avoid long polished speeches.
- A useful principle:
- `subtext > literal romantic wording`
- If the dialogue does not comfortably fit the duration, remove lines rather than making the characters speak unnaturally fast.
- ## Micro-expression rules
- Translate abstract emotions into visible behavior.
- Do not write only:
- `she gets angry` or `she becomes happy`.
- Instead describe sequences such as:
- - smile remains for a beat
- - she realizes what he said
- - the smile fades slightly
- - eyes narrow a little
- - lips press together
- - one corner of the mouth almost rises
- - she looks away
- - secretly glances back
- - tries not to laugh
- - lowers her head
- - shoulders move with suppressed laughter
- - finally breaks into a genuine smile
- For playful fake anger, emphasize the "trying not to laugh" progression.
- Avoid influencer-style exaggerated faces.
- ## Eye-line rules
- Do not make the lead stare into the lens continuously.
- Use changing eye behavior:
- `look at partner → look away → lower gaze → glance back → pretend to stay annoyed → suppress smile → re-establish eye contact`
- The eye-line must support the emotional beat.
- ## POV and camera
- Default to realistic close-distance relationship POV when appropriate.
- For boyfriend POV:
- - the boyfriend's face must remain off-screen
- - a hand, forearm, sleeve edge, or blurred shoulder may appear
- - never reveal the full male body unless requested
- - treat the female lead looking into camera as looking at her partner
- Use mild handheld breathing motion.
- Avoid:
- - aggressive push-ins
- - rapid zooms
- - unnecessary orbit shots
- - music-video camera movement
- - a cut for every dialogue line
- Recommended shot counts:
- - 10s: 1–2 shots
- - 15s: 1–3 shots
- - 30s: 3–4 longer shots
- - 60s: usually split into two 30s parts if the target model has a 30s generation limit
- ## Mirror rule
- If a mirror exists, explicitly prevent accidental reflections of:
- - partner's face
- - camera operator
- - camera / phone / rig
- - crew
- - third person
- Mirror geometry should remain consistent with the room.
- ## Physical interaction
- Use only a few simple, readable actions.
- Examples:
- - light cheek pinch
- - move a strand of hair
- - brief hand touch
- - push a hand away
- - tug a sleeve
- - touch nose tip
- - grab a pillow
- - block the camera
- - lean closer
- - short hug
- Interactions should be brief, natural, and physically plausible.
- For hand contact, include:
- - correct finger count
- - stable hand anatomy
- - no fusion
- - no extra hands
- - no penetration
- - real contact
- - natural entry and exit from frame
- - no sudden disappearance
- ## Story engine
- For light couple scenes, prefer:
- `normal conversation → tiny misunderstanding / tease → small emotional reaction → awkward attempt to fix it → unintended funny line → suppressed laughter → small reversal → genuine reaction`
- Possible everyday hooks:
- - one person talks too much
- - blanket stealing
- - room temperature
- - forgotten reply
- - ugly photo
- - snacks
- - pillow fight
- - wrong nickname
- - gaming result
- - who turns off the light
- - who gets water
- - messy hair
- - one person refusing to admit they are smiling
- Avoid major betrayal, breakup, death, or third-party drama unless explicitly requested.
- ## Duration logic
- ### Under 10 seconds
- Use:
- - one setup
- - one interaction
- - one payoff
- ### 15 seconds
- Typical structure:
- - 00:00–00:04 setup
- - 00:04–00:09 tiny conflict
- - 00:09–00:13 interaction
- - 00:13–00:15 payoff
- Usually 4–7 short dialogue lines maximum.
- ### 30 seconds
- Typical structure:
- - 00:00–00:06 natural setup
- - 00:06–00:13 conflict develops
- - 00:13–00:19 physical interaction
- - 00:19–00:25 comic escalation
- - 00:25–00:30 reversal / laugh / ending beat
- Usually 8–15 short dialogue lines, depending on pacing.
- ### 60 seconds
- If the chosen model is practically limited to ~30 seconds per generation, produce:
- - Part 1: 00:00–00:30
- - Part 2: 00:30–01:00
- Each part should have a mini-payoff and a clean continuity hook.
- ## Audio
- Unless requested otherwise for natural daily-life scenes:
- - no background music
- - no subtitles
- - no text overlays
- - no narration
- Keep realistic diegetic sound:
- - room tone
- - air conditioner
- - bedding friction
- - clothing friction
- - hair movement
- - breathing
- - suppressed laughter
- - natural speech
- Leave small pauses between lines.
- ## Model adaptation
- ### Seedance
- Prioritize:
- - clean continuous actions
- - explicit time ranges
- - moderate dialogue density
- - limited complex hand choreography
- - clear cause-and-effect blocking
- ### Hailuo
- Prioritize:
- - strong identity lock
- - explicit action descriptions
- - one main action at a time
- - clear continuity reminders
- ### Kling
- Prioritize:
- - spatial clarity
- - camera-to-subject relation
- - consistent blocking
- - slightly more cinematic camera language when useful
- ### Veo / Sora
- May include richer:
- - sound cues
- - environmental details
- - lens / camera behavior
- - dialogue timing
- - subtle performance direction
- Do not add technical camera jargon unless it improves execution.
- ## Output format
- Produce a clean, directly copyable prompt with these sections when relevant:
- 1. Character Consistency
- 2. Outfit Consistency
- 3. Core Setting
- 4. Scene
- 5. Camera / POV
- 6. Timeline
- 7. Performance Notes
- 8. Male / Off-screen Partner Performance
- 9. Continuity
- 10. Sound Design
- 11. Negative Constraints
- 12. Model-specific notes
- For each timeline block, include as needed:
- - action
- - micro-expression
- - eye-line
- - dialogue
- - pause
- - camera behavior
- Do not mechanically include empty sections.
- ## Editing an existing user prompt
- If the user already wrote a detailed prompt, preserve the core story.
- Apply this order:
- `preserve story → reduce redundant dialogue → add pauses → add eye-line changes → add micro-expressions → fix physical continuity → reduce melodrama → improve relationship familiarity → simplify actions that are difficult for video models`
- Do not overwrite distinctive user ideas unless they conflict with the user's stated constraints.
- ## Negative constraints
- Use relevant items, not necessarily every line:
- - No face identity drift.
- - No face swapping.
- - No hairstyle changes.
- - No outfit changes.
- - No sudden makeup changes.
- - No unexplained age changes.
- - No extra people.
- - No visible camera operator.
- - No visible recording equipment.
- - No partner face in mirrors when POV requires them to remain unseen.
- - No extra hands.
- - No malformed fingers.
- - No hand penetration.
- - No disappearing hands.
- - No exaggerated facial acting.
- - No influencer-style acting.
- - No idol-drama acting.
- - No melodramatic romance.
- - No excessive blinking.
- - No fixed staring.
- - No subtitles unless requested.
- - No text unless requested.
- - No background music unless requested.
- - No random scene changes.
- - No jumpy continuity.
- ## Final quality check
- Before returning the prompt, verify:
- - one stable identity per character
- - adult characters when romance is involved
- - outfit continuity
- - location continuity
- - dialogue fits duration
- - pauses remain possible
- - POV rules are respected
- - mirror reflections are safe
- - hand actions are feasible
- - no unnecessary extra characters
- - emotions are expressed through visible micro-actions
- - ending has a readable final beat
- The final result should feel less like a scripted romance scene and more like a believable moment that happened to be recorded.
README.md · README · 338 lines
View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)
- # Reference Video Director
- <p align="center">
- <a href="./README.md"><img alt="中文" src="https://img.shields.io/badge/语言-中文-111111"></a>
- <a href="./README_EN.md"><img alt="English" src="https://img.shields.io/badge/Language-English-111111"></a>
- </p>
- **Reference Video Director** 是一个面向参考图驱动视频创作的 Agent Skill。
- 它会根据角色参考图、服装参考图、场景参考图、人物关系、剧情想法和目标时长,生成更适合视频模型直接执行的完整提示词,重点处理人物一致性、情侣 POV、自然对白、微表情、手部互动、镜头连续性与真实生活感。
- # 🚀 新手 3 分钟上手
- 完全没用过 Skill 也没关系,按下面做就行。
- ## 第一步:下载
- ### 方法 A:直接下载 ZIP(最适合新手)
- 打开本仓库后,点击:
- ```text
- Code → Download ZIP
- ```
- 下载完成后解压,得到 `reference-video-director` 文件夹。
- **不要只下载 `SKILL.md`,建议保留整个文件夹结构。**
- ### 方法 B:Git Clone
- 会用 Git 的话可以直接:
- ```bash
- git clone https://github.com/sgyno09-source/reference-video-director.git
- ```
- ## 第二步:安装 / 导入
- 这个仓库本身就是一个 Skill 目录,核心入口是:
- ```text
- SKILL.md
- ```
- 如果你的 Agent、CLI、Codex 或其他客户端支持 Skills:
- 1. 找到它的 Skills / Agent Skills / Custom Skills 导入入口。
- 2. 选择刚才下载并解压的 `reference-video-director` 文件夹。
- 3. 保持 `SKILL.md` 和 `references/` 等文件原有位置不变。
- 4. 导入完成后重新打开会话或刷新 Skill 列表。
- 不同客户端的安装入口可能不同,但原则一样:**导入整个 Skill 文件夹,而不是把里面的文字复制出来。**
- ## 第三步:开始使用
- 安装后不需要背复杂命令。
- 上传参考图,然后直接告诉它:
- ```text
- 参考这张女生图
- 做一个30秒情侣POV
- 保持同一个人物和发型
- 服装不要变化
- 场景是晚上卧室
- 女生性格俏皮爱笑
- 男友不露脸
- 剧情:
- 男友说错一句话
- 女生先愣一下
- 然后假装生气
- 最后又被逗笑
- 模型:Seedance
- 对白自然一点
- 不要字幕
- ```
- Skill 会帮你自动补全:
- ```text
- 参考图绑定
- 人物与服装一致性
- 场景与光线
- Camera / POV
- 分秒时间轴
- 动作与微表情
- 对白与停顿
- 手部互动
- 声音设计
- 连续性控制
- 负面约束
- 模型适配
- ```
- 你负责给**参考图 + 大概剧情**,剩下的交给 Skill。
- ## 多张参考图怎么写
- 如果同时上传人物、服装、场景参考图,最好明确告诉它每张图负责什么:
- ```text
- Mixed 1:女主人物参考
- Mixed 2:服装参考
- Mixed 3:卧室场景参考
- 使用 Mixed 1 作为唯一女主角
- 服装严格参考 Mixed 2
- 场景结构参考 Mixed 3
- 生成30秒情侣POV
- ```
- 这样可以减少换脸、换衣服、场景漂移等问题。
- ## 只会一句剧情也可以
- 不需要自己会写分镜。
- 例如直接说:
- ```text
- 参考这个女生
- 做一个30秒视频
- 男朋友说错一句话
- 她从开心到愣住
- 再到假装生气
- 最后被逗笑
- 帮我完整设计
- ```
- Skill 会自动把一句话扩展成可执行的视频提示词。
- ## 做 15 秒 / 30 秒视频
- 直接告诉它时长即可:
- ```text
- 做15秒,节奏快一点,只保留一个转折
- ```
- 或者:
- ```text
- 做30秒,保留起因、假生气、哄人和最后笑出来
- ```
- Skill 会根据时长自动控制对白和动作密度,不会把一分钟剧情硬塞进 30 秒。
- ## 做两个连续的 30 秒
- 如果模型单次只能生成 30 秒,可以这样说:
- ```text
- 这段剧情一共60秒
- 拆成两个30秒
- 第二段必须接第一段最后一个动作
- 人物、服装、发型、场景、灯光和情绪状态保持一致
- ```
- Skill 会为第二段补上上一段结束时的人物位置、动作和状态,方便连续生成。
- ## POV / 男友不露脸怎么写
- 例如:
- ```text
- 男友使用第一人称POV
- 全程不露脸
- 只允许手、前臂、肩部或画外音进入画面
- 镜子里不要出现男友和摄影设备
- ```
- 这类空间关系、镜面和手部互动也是本 Skill 重点处理的部分。
- ## 指定视频模型
- 可以直接写模型名称:
- ```text
- 模型:Seedance
- ```
- ```text
- 模型:Hailuo
- ```
- ```text
- 模型:Sora
- ```
- Skill 会根据模型调整动作密度、时间轴、镜头与连续性描述。没有指定模型时,会输出通用版本。
- ## 一个推荐的新手模板
- 不会写需求时,复制下面这段改内容就行:
- ```text
- 参考图:
- 图1是人物
- 图2是服装
- 图3是场景
- 时长:30秒
- 模型:Seedance
- 画幅:9:16
- 人物关系:情侣
- 女生性格:俏皮、爱笑、情绪写在脸上
- 男友:不露脸,第一人称POV
- 剧情:
- 这里写一句你想发生的事情
- 要求:
- 保持同一人物
- 保持同一服装
- 对白自然
- 加强眼神和微表情
- 动作不要太密
- 不要字幕
- 不要突然换场景
- ```
- 把最后的剧情换成自己的想法,就可以开始用了。
- # 适合什么
- - 情侣 POV / 恋爱短剧
- - 女友假生气 / 男友哄人
- - 居家、卧室、旅行、餐厅等生活化场景
- - 参考图角色一致性
- - 15 秒 / 30 秒短剧情
- - Seedance、Hailuo、Kling、Veo、Sora 等视频模型
- - 将粗略剧情改造成时间轴 Prompt
- - 将“太像演戏”的对白改得更自然
- ## 核心原则
- **真实感 > 戏剧性**
- **人物反应 > 漂亮台词**
- **情侣默契 > 强行暧昧**
- **微表情 > 大动作**
- **停顿 > 连续念台词**
- **小事情 > 大冲突**
- ## 输入示例
- ```text
- 使用 {{Mixed 1}} 作为唯一女主角参考,
- 服装参考 {{Mixed 2}}。
- 做一个 30 秒男友 POV 情侣短剧:
- 女友本来在聊天,男友说“你今天话好多”,
- 女友开始假生气,最后被男友一句话逗笑。
- 模型:Seedance
- ```
- Skill 会自动完成:
- - Character Consistency
- - Outfit Consistency
- - Core Setting
- - Scene
- - Camera / POV
- - Timeline
- - Dialogue
- - Micro-expressions
- - Physical Interaction
- - Continuity
- - Sound Design
- - Negative Constraints
- - Model Adaptation
- ## 安装
- 这个仓库本身就是一个 Skill 目录,核心入口文件是:
- ```text
- SKILL.md
- ```
- 兼容 Agent Skills / Codex Skill 风格的客户端,可直接使用仓库目录作为 Skill 来源。
- ## 目录
- ```text
- reference-video-director/
- ├── SKILL.md
- ├── README.md
- ├── README_EN.md
- └── references/
- ├── zh-CN.md
- └── en.md
- ```
- ## 语言
- Skill 会自动跟随用户当前使用的语言。
- - 中文输入 → 中文输出
- - English input → English output
- - 用户明确指定语言 → 按指定语言输出
- ## RISE 火焰电影提示词
- 仓库新增独立技能 [RISE 火焰电影提示词](skills/create-rise-fire-video/SKILL.md),可单独导入 `skills/create-rise-fire-video/` 整个目录,也可以通过主技能自动调用。
- 默认方案为30秒、16:9连续镜头:点烟、吐烟、抛出同一根火柴、落地后火势沿RISE路径扩展、镜头持续后拉、22秒后翘二郎腿,最终完整人物、椅子、双脚与RISE互不遮挡。支持换人物、换文字和调整时长。
- 人物参考图需由使用者提供。技能输出视频提示词,不会自行生成视频。
- 调用示例:
- ```text
- 使用 create-rise-fire-video
- 图1作为唯一人物参考
- 做30秒RISE火焰视频完整提示词
- 保持全身、椅子和字母完整无遮挡
- ```
- 完整时间轴与约束见 [30秒方案](skills/create-rise-fire-video/references/rise-30s.md)。
- ---
- Made for reference-driven video prompt workflows.
Ratings and comments
Sign in to rate this skill.
Sign inFree to use. If it’s helped, pass it on.